LLM: Enable BigDL IPEX optimization for int4 (#10319)

Enable BigDL IPEX optimization for int4

python/llm/dev/benchmark/all-in-one/config.yaml

@@ -18,6 +18,8 @@ test_api:
   - "optimize_model"
   - "pytorch_autocast_bf16"
   # - "transformer_autocast_bf16"
+  # - "bigdl_ipex_bf16"
+  # - "bigdl_ipex_int4"
   # - "ipex_fp16_gpu" # on Intel GPU
   # - "bigdl_fp16_gpu" # on Intel GPU
   # - "transformer_int4_gpu"  # on Intel GPU

python/llm/dev/benchmark/all-in-one/run.py

@@ -92,6 +92,10 @@ def run_model(repo_id, test_api, in_out_pairs, local_model_hub=None, warm_up=1,
         result = run_transformer_int4_loadlowbit_gpu_win(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials, num_beams, low_bit, cpu_embedding, batch_size, streaming)
     elif test_api == 'transformer_autocast_bf16':
         result = run_transformer_autocast_bf16(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials, num_beams, batch_size)
+    elif test_api == 'bigdl_ipex_bf16':
+        result = run_bigdl_ipex_bf16(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials, num_beams, batch_size)
+    elif test_api == 'bigdl_ipex_int4':
+        result = run_bigdl_ipex_int4(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials, num_beams, batch_size)
     elif test_api == 'deepspeed_optimize_model_gpu':
         result = run_deepspeed_optimize_model_gpu(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials, num_beams, low_bit, batch_size)
 
@@ -1079,6 +1083,148 @@ def run_transformer_autocast_bf16( repo_id,
                                           actual_in_len, actual_out_len, load_time])
     return result
 
+
+def run_bigdl_ipex_bf16(repo_id,
+                    local_model_hub,
+                    in_out_pairs,
+                    warm_up,
+                    num_trials,
+                    num_beams,
+                    batch_size):
+    from bigdl.llm.transformers import AutoModel, AutoModelForCausalLM
+    from transformers import AutoTokenizer, LlamaTokenizer
+
+    os.environ["BIGDL_OPT_IPEX"] = "true"
+
+    model_path = get_model_path(repo_id, local_model_hub)
+    # Load model in bf16,
+    # which convert the relevant layers in the model into BF16 format
+    st = time.perf_counter()
+    if repo_id in CHATGLM_IDS:
+        model = AutoModel.from_pretrained(model_path, load_in_low_bit='bf16', trust_remote_code=True, torch_dtype=torch.bfloat16,
+                                          use_cache=True, torchscript=True)
+        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    elif repo_id in LLAMA_IDS:
+        model = AutoModelForCausalLM.from_pretrained(model_path, load_in_low_bit='bf16', trust_remote_code=True, torch_dtype=torch.bfloat16,
+                                                     use_cache=True, torchscript=True)
+        tokenizer = LlamaTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    else:
+        model = AutoModelForCausalLM.from_pretrained(model_path, load_in_low_bit='bf16', trust_remote_code=True, torch_dtype=torch.bfloat16,
+                                                     use_cache=True, torchscript=True)
+        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    if not hasattr(model.config, "token_latency"):
+        model.config.token_latency = True
+    end = time.perf_counter()
+    load_time = end - st
+    print(">> loading of model costs {}s".format(load_time))
+
+    result = {}
+    with torch.inference_mode(), torch.autocast("cpu"):
+        for in_out in in_out_pairs:
+            in_out_len = in_out.split("-")
+            in_len = int(in_out_len[0])
+            out_len = int(in_out_len[1])
+            # As different tokenizer has different encodings,
+            # in_len.txt maybe shorter than we need,
+            # use much longer context to make sure input length
+            test_length = min(in_len*2, 8192)
+            while test_length not in [32, 256, 1024, 2048, 8192]:
+                test_length = test_length * 2
+            input_str = open(f"prompt/{test_length}.txt", 'r').read()
+            # As different tokenizer has different encodings,
+            # slice the input_ids to ensure the prompt length is required length.
+            input_ids = tokenizer.encode(input_str, return_tensors="pt")
+            input_ids = input_ids[:, :in_len]
+            true_str = tokenizer.batch_decode(input_ids)[0]
+            input_list = [true_str] * batch_size
+            input_ids = tokenizer(input_list, return_tensors="pt").input_ids
+            actual_in_len = input_ids.shape[1]
+            result[in_out] = []
+            for i in range(num_trials + warm_up):
+                st = time.perf_counter()
+                output_ids, total_list = model.generate(input_ids, do_sample=False, max_new_tokens=out_len,
+                                            num_beams=num_beams)
+                end = time.perf_counter()
+                print("model generate cost: " + str(end - st))
+                output = tokenizer.batch_decode(output_ids)
+                print(output[0])
+                actual_out_len = output_ids.shape[1] - actual_in_len
+                if i >= warm_up:
+                    result[in_out].append([total_list[0], np.mean(total_list[1:]), 0,
+                                          actual_in_len, actual_out_len, load_time])
+    return result
+
+
+def run_bigdl_ipex_int4(repo_id,
+                    local_model_hub,
+                    in_out_pairs,
+                    warm_up,
+                    num_trials,
+                    num_beams,
+                    batch_size):
+    from bigdl.llm.transformers import AutoModel, AutoModelForCausalLM
+    from transformers import AutoTokenizer, LlamaTokenizer
+
+    os.environ["BIGDL_OPT_IPEX"] = "true"
+
+    model_path = get_model_path(repo_id, local_model_hub)
+
+    st = time.perf_counter()
+    if repo_id in CHATGLM_IDS:
+        model = AutoModel.from_pretrained(model_path, load_in_low_bit='sym_int4', trust_remote_code=True, torch_dtype='auto',
+                                          use_cache=True, torchscript=True)
+        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    elif repo_id in LLAMA_IDS:
+        model = AutoModelForCausalLM.from_pretrained(model_path, load_in_low_bit='sym_int4', trust_remote_code=True, torch_dtype='auto',
+                                                     use_cache=True, torchscript=True)
+        tokenizer = LlamaTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    else:
+        model = AutoModelForCausalLM.from_pretrained(model_path, load_in_low_bit='sym_int4', trust_remote_code=True, torch_dtype='auto',
+                                                     use_cache=True, torchscript=True)
+        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    if not hasattr(model.config, "token_latency"):
+        model.config.token_latency = True
+    end = time.perf_counter()
+    load_time = end - st
+    print(">> loading of model costs {}s".format(load_time))
+
+    result = {}
+    with torch.inference_mode(), torch.autocast("cpu"):
+        for in_out in in_out_pairs:
+            in_out_len = in_out.split("-")
+            in_len = int(in_out_len[0])
+            out_len = int(in_out_len[1])
+            # As different tokenizer has different encodings,
+            # in_len.txt maybe shorter than we need,
+            # use much longer context to make sure input length
+            test_length = min(in_len*2, 8192)
+            while test_length not in [32, 256, 1024, 2048, 8192]:
+                test_length = test_length * 2
+            input_str = open(f"prompt/{test_length}.txt", 'r').read()
+            # As different tokenizer has different encodings,
+            # slice the input_ids to ensure the prompt length is required length.
+            input_ids = tokenizer.encode(input_str, return_tensors="pt")
+            input_ids = input_ids[:, :in_len]
+            true_str = tokenizer.batch_decode(input_ids)[0]
+            input_list = [true_str] * batch_size
+            input_ids = tokenizer(input_list, return_tensors="pt").input_ids
+            actual_in_len = input_ids.shape[1]
+            result[in_out] = []
+            for i in range(num_trials + warm_up):
+                st = time.perf_counter()
+                output_ids, total_list = model.generate(input_ids, do_sample=False, max_new_tokens=out_len,
+                                            num_beams=num_beams)
+                end = time.perf_counter()
+                print("model generate cost: " + str(end - st))
+                output = tokenizer.batch_decode(output_ids)
+                print(output[0])
+                actual_out_len = output_ids.shape[1] - actual_in_len
+                if i >= warm_up:
+                    result[in_out].append([total_list[0], np.mean(total_list[1:]), 0,
+                                          actual_in_len, actual_out_len, load_time])
+    return result
+
+
 def run_deepspeed_optimize_model_gpu(repo_id,
                                      local_model_hub,
                                      in_out_pairs,
@@ -1192,6 +1338,7 @@ def run_deepspeed_optimize_model_gpu(repo_id,
     torch.xpu.empty_cache()
     return result
 
+
 if __name__ == '__main__':
     from omegaconf import OmegaConf
     conf = OmegaConf.load(f'{current_dir}/config.yaml')

python/llm/src/bigdl/llm/transformers/convert.py

@@ -611,13 +611,12 @@ def ggml_convert_low_bit(model, qtype, optimize_model=True,
     modules_to_not_convert = [] if modules_to_not_convert is None else modules_to_not_convert
 
     # using ipex optimizer before changing to bigdl linear
-    _enable_ipex = os.getenv("BIGDL_OPT_IPEX")
+    _enable_ipex = get_enable_ipex()
-    _enable_ipex = (_enable_ipex is not None) and (_enable_ipex.lower() == "true")
+
-    _enable_ipex = _enable_ipex and (qtype == ggml_tensor_qtype["bf16"])
+    if device == "cpu":
-    if (device == "cpu") and (qtype == ggml_tensor_qtype["bf16"]):
         logger.info(f"BIGDL_OPT_IPEX: {_enable_ipex}")
     if _enable_ipex:
-        model = _optimize_ipex(model)
+        model = _optimize_ipex(model, qtype)
         return model
 
     if optimize_model:
@@ -686,12 +685,19 @@ def replace_func(m, target_m, func_name, new_func):
         replace_func(sub_m, target_m, func_name, new_func)
 
 
-def _optimize_ipex(model):
+def get_enable_ipex():
+    _enable_ipex = os.getenv("BIGDL_OPT_IPEX")
+    _enable_ipex = (_enable_ipex is not None) and (_enable_ipex.lower() == "true")
+    return _enable_ipex
+
+
+def _optimize_ipex(model, qtype=ggml_tensor_qtype["bf16"]):
+    import intel_extension_for_pytorch as ipex
     from intel_extension_for_pytorch.transformers.optimize import model_convert_reference
     from transformers.modeling_attn_mask_utils import AttentionMaskConverter
     from bigdl.llm.transformers.convert_ipex import (
         _ipex_optimize_model, _ipex_jit, _make_causal_mask,
-        _llama_model_forward_4_35, convert_function, GLM_get_masks
+        _llama_model_forward_4_35, convert_function, GLM_get_masks,
     )
 
     model = model_convert_reference(model)
@@ -718,7 +724,7 @@ def _optimize_ipex(model):
         # baichuan2
         rms_classes.append(type(model.model.layers[0].input_layernorm))
 
-    _ipex_optimize_model(model, rms_classes)
+    model = _ipex_optimize_model(model, rms_classes, qtype)
     return _ipex_jit(model)
 
 

python/llm/src/bigdl/llm/transformers/convert_ipex.py

@@ -44,10 +44,14 @@ from intel_extension_for_pytorch.transformers.optimize import (
 from intel_extension_for_pytorch.cpu._auto_kernel_selection import (
     _enable_tpp,
     _using_tpp,
+    _disable_tpp
 )
+from bigdl.llm.ggml.quantize import ggml_tensor_qtype
+from bigdl.llm.transformers.convert import get_enable_ipex
+import os
 
 
-def _ipex_optimize_rmsnorm(_model, supported_classes):
+def _ipex_optimize_rmsnorm(_model, supported_classes, is_tpp=False, is_woq=False):
     from intel_extension_for_pytorch.transformers.models.cpu.fusions.mha_fusion import _IPEXRMSNorm
     for supported_class in supported_classes:
         lowering_class_cpu(
@@ -55,12 +59,12 @@ def _ipex_optimize_rmsnorm(_model, supported_classes):
             supported_class,
             _IPEXRMSNorm,
             _model.config,
-            tpp=False,
+            tpp=is_tpp,
-            woq=False,
+            woq=is_woq,
         )
 
 
-def _ipex_optimize_decoder(model):
+def _ipex_optimize_decoder(model, is_tpp=False, is_woq=False):
     from intel_extension_for_pytorch.transformers.models.reference.modules.decoder import (
         _IPEXDecoderLayerRef
     )
@@ -73,12 +77,12 @@ def _ipex_optimize_decoder(model):
             supported_mlp_class,
             _IPEXDecoderLayerCPU,
             model.config,
-            tpp=True if _using_tpp() else False,
+            tpp=is_tpp,
-            woq=False,
+            woq=is_woq,
         )
 
 
-def _ipex_optimize_attention(model):
+def _ipex_optimize_attention(model, is_tpp=False, is_woq=False):
     from intel_extension_for_pytorch.transformers.models.reference.modules.attentions import (
         _IPEXAttentionRef
     )
@@ -91,18 +95,47 @@ def _ipex_optimize_attention(model):
             supported_mha_class,
             _IPEXAttentionCPU,
             model.config,
-            tpp=True if _using_tpp() else False,
+            tpp=is_tpp,
-            woq=False,
+            woq=is_woq,
         )
 
 
-def _ipex_optimize_model(model, rms_classes):
+def _ipex_optimize_model(model, rms_classes, qtype):
-    _enable_tpp()
     import intel_extension_for_pytorch as ipex
-    ipex.optimize(model.eval(), dtype=torch.bfloat16, inplace=True).eval()
+    from intel_extension_for_pytorch.transformers.models.reference.models import output_hook
-    _ipex_optimize_rmsnorm(model, rms_classes)
+    from intel_extension_for_pytorch.transformers.optimize import ipex_quantization_flow
-    _ipex_optimize_attention(model)
+
-    _ipex_optimize_decoder(model)
+    is_woq = False
+    is_quantization = False
+    _disable_tpp()
+    if qtype == ggml_tensor_qtype["bf16"]:
+        _enable_tpp()
+        model = ipex.optimize(model.eval(), dtype=torch.bfloat16, inplace=True).eval()
+    elif qtype == ggml_tensor_qtype["sym_int4"]:
+        is_quantization = True
+        is_woq = True
+        act_quant_mode_dict = {
+            "PER_TENSOR": ipex.quantization.WoqActQuantMode.PER_TENSOR,
+            "PER_IC_BLOCK": ipex.quantization.WoqActQuantMode.PER_IC_BLOCK,
+            "PER_BATCH": ipex.quantization.WoqActQuantMode.PER_BATCH,
+            "PER_BATCH_IC_BLOCK": ipex.quantization.WoqActQuantMode.PER_BATCH_IC_BLOCK,
+        }
+        qconfig = ipex.quantization.get_weight_only_quant_qconfig_mapping(
+            weight_dtype=torch.quint4x2,  # INT4
+            lowp_mode=ipex.quantization.WoqLowpMode.INT8,
+            act_quant_mode=act_quant_mode_dict["PER_IC_BLOCK"],
+            group_size=-1,
+        )
+        model = ipex_quantization_flow(model, torch.bfloat16, None, qconfig, None)
+
+    is_tpp = _using_tpp()
+
+    _ipex_optimize_rmsnorm(model, rms_classes, is_tpp=is_tpp, is_woq=is_woq)
+    _ipex_optimize_attention(model, is_tpp=is_tpp, is_woq=is_woq)
+    _ipex_optimize_decoder(model, is_tpp=is_tpp, is_woq=is_woq)
+
+    model.register_forward_hook(output_hook, with_kwargs=True)
+    return model
 
 
 def _ipex_jit(model):
@@ -152,9 +185,7 @@ def GLM_get_masks(self, input_ids, past_key_values, padding_mask=None):
         else:  # discrete kv cache
             past_length = past_key_values[0][0].shape[-2]
 
-    import os
+    _enable_ipex = get_enable_ipex()
-    _enable_ipex = os.getenv("BIGDL_OPT_IPEX")
-    _enable_ipex = (_enable_ipex is not None) and (_enable_ipex.lower() == "true")
     # always call for jit
     if past_length or _enable_ipex:
         full_attention_mask = torch.cat(
@@ -191,9 +222,7 @@ def _make_causal_mask(
     mask_cond = torch.arange(mask.size(-1), device=device)
     mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
 
-    import os
+    _enable_ipex = get_enable_ipex()
-    _enable_ipex = os.getenv("BIGDL_OPT_IPEX")
-    _enable_ipex = (_enable_ipex is not None) and (_enable_ipex.lower() == "true")
     if _enable_ipex or past_key_values_length > 0:
         mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)  # noqa
 

python/llm/src/bigdl/llm/transformers/speculative.py

@@ -30,6 +30,7 @@ from typing import TYPE_CHECKING, Any, Callable, Dict, List, Optional, Tuple, Un
 from transformers import top_k_top_p_filtering, GenerationConfig, \
     LogitsProcessorList, StoppingCriteriaList
 from bigdl.llm.utils.common import invalidInputError
+from transformers.modeling_outputs import CausalLMOutputWithPast
 
 # patch GenerationMixin.generate
 from transformers import GenerationMixin
@@ -533,15 +534,16 @@ def speculative_generate(self,
     past_key_values = None
     past_key_values_storage = []
 
-    _enable_ipex = os.getenv("BIGDL_OPT_IPEX")
+    from bigdl.llm.transformers.convert import get_enable_ipex
-    _enable_ipex = (_enable_ipex is not None) and (_enable_ipex.lower() == "true")
+    _enable_ipex = get_enable_ipex()
+
     if _enable_ipex:
         if not ((self.config.model_type == 'baichuan') or
                 ('llama' in self.config.model_type) or
                 ("mistral" in self.config.model_type) or
                 ("qwen" in self.config.model_type) or
                 ("chatglm" in self.config.model_type)):
-            invalidInputError(False, "BigDL Speculative Decoding with IPEX BF16 only supports \
+            invalidInputError(False, "BigDL Speculative Decoding with IPEX only supports \
                               Llama, Baichuan2, Mistral, ChatGLM and Qwen models currently.")
         if "chatglm" in self.config.model_type:
             global query_group_size
@@ -579,6 +581,11 @@ def speculative_generate(self,
                           attention_mask=attention_mask,
                           return_dict=True,
                           use_cache=True)
+            if _enable_ipex:
+                output = CausalLMOutputWithPast(
+                    logits=output[0],
+                    past_key_values=output[1],
+                )
             logits = output['logits']
             logits = logits[:, -1:]
             logits[:, -1, :] = logits_processor(current_input_ids, logits[:, -1, :])
@@ -602,7 +609,7 @@ def speculative_generate(self,
             draft_current_input_ids = current_input_ids
             # Target model KV cache to draft model
 
-            if self.device.type == 'cpu':
+            if self.device.type == 'cpu' and (not _enable_ipex):
                 # init past_key_values_storage and assign initial fp32 value
                 if step == 1:
                     past_key_values_storage = \
@@ -652,7 +659,57 @@ def speculative_generate(self,
                     past_length = draft_past_key_values[0][0].size(2)
                     position_ids = torch.Tensor([[past_length]]).long().to(self.device)
                     forward_args["position_ids"] = position_ids
-                draft_output = draft_model(**forward_args)
+
+                if _enable_ipex:
+                    if any(keyword in self.config.model_type
+                            for keyword in ["llama", "chatglm", "mistral"]):
+                        past_key_value_len = draft_past_key_values[0][0].shape[2]
+                        position_ids = torch.Tensor([[past_key_value_len + step_draft]]).long()
+                        position_ids = position_ids[:, :-draft_current_input_ids.size(0)]
+                        draft_output = draft_model.trace_graph(
+                            input_ids=draft_current_input_ids,
+                            attention_mask=draft_attention_mask,
+                            position_ids=position_ids,
+                            past_key_values=draft_past_key_values,
+                        )
+                    elif self.config.model_type == "baichuan":
+                        if self.config.hidden_size == 4096:
+                            past_key_value_len = draft_past_key_values[0][0].shape[2]
+                            seq_len = draft_current_input_ids.shape[1]
+                            seq_len_with_past = seq_len + past_key_value_len
+                            position_ids = torch.arange(past_key_value_len,
+                                                        seq_len_with_past,
+                                                        dtype=torch.long,
+                                                        device=draft_current_input_ids.device)
+                            position_ids = position_ids.unsqueeze(0).view(-1, seq_len)
+                            draft_output = draft_model.trace_graph(
+                                input_ids=draft_current_input_ids,
+                                attention_mask=draft_attention_mask,
+                                position_ids=position_ids,
+                                past_key_values=draft_past_key_values,
+                            )
+                        elif self.config.hidden_size == 5120:
+                            draft_output = draft_model.trace_graph(
+                                input_ids=draft_current_input_ids,
+                                attention_mask=draft_attention_mask,
+                                past_key_values=draft_past_key_values,
+                            )
+                    elif "qwen" in self.config.model_type:
+                        draft_output = draft_model.trace_graph(
+                            input_ids=draft_current_input_ids,
+                            attention_mask=draft_attention_mask,
+                            past_key_values=draft_past_key_values,
+                        )
+                    else:
+                        invalidInputError(False, "BigDL Speculative Decoding with IPEX only supports \
+                              Llama, Baichuan2, Mistral, ChatGLM and Qwen models currently.")
+
+                    draft_output = CausalLMOutputWithPast(
+                        logits=draft_output[0],
+                        past_key_values=draft_output[1],
+                    )
+                else:
+                    draft_output = draft_model(**forward_args)
                 temp_input_ids = torch.cat((input_ids, generate_ids[:, :step],
                                             draft_generate_ids[:, 1:step_draft+1]), dim=-1)
                 logits = draft_output['logits']
@@ -848,7 +905,6 @@ def speculative_generate(self,
             # Clean up target model KV cache
             if max_of_max_matched != max_matched:
                 output_ids = output_ids[:, :max_matched]
-                # For Qwen
                 if _enable_ipex:
                     cur_len = past_key_values[0][0].size(1)
                     delta = max_of_max_matched - max_matched
@@ -890,7 +946,7 @@ def speculative_generate(self,
                         ]
 
             # Each iter assign new_matched kv_cache to past_key_values1
-            if self.device.type == 'cpu':
+            if self.device.type == 'cpu' and (not _enable_ipex):
                 _update_past_key_values_storage_cpu(self, past_key_values, past_key_values_storage,
                                                     original_draft_past_key_values,
                                                     _enable_ipex)